FREE FLIGHT WIND TUNNEL TESTS FOR PARAMETER IDENTIFICATION

Page created by Roland Armstrong
 
CONTINUE READING
FREE FLIGHT WIND TUNNEL TESTS FOR
                          PARAMETER IDENTIFICATION
                                  Jan Nowack and Wolfgang Alles

                                            Chair of Flight Dynamics
                                           RWTH Aachen University
                                           D-52062 Aachen, Germany

                                           Received: November 12, 2008.

                                                ABSTRACT
The Chair of Flight Dynamics at the RWTH Aachen University is conducting research on a method for
identification of flight mechanical characteristics on free flying models in a wind tunnel. The main goal is to
create a reproducible free flight environment for cost effective identification of important values even in an early
design stage. The method will combine the advantages of free flight with wind tunnel techniques as it takes the
free flight into a reproducible environment under laboratory conditions. The paper gives an overview of the
project and provides insight into the work done so far.

Keywords: Identification, simulation, nonlinear dynamic inversion, Pseudo Control Hedging, MATLAB,
Simulink, dSPACE, wind tunnel

                                           1 INTRODUCTION
   Aircraft are characterized by numerous closely coupled subsystems. A separate
design of the single items is not possible and experts' knowledge of each item is
needed. First of all this affects the aerodynamics and flight mechanical characteristics.
Hence, the creation of an authentic aerodynamic and flight mechanical dataset can only
be carried out with high costs.
   Typical methods for the creation of an aerodynamic and flight mechanical dataset
consist of theoretical and experimental techniques. Even if the further development of
the numerical methods delivers deeper insights into the fluid mechanics, the
application on complex configurations or in an early design process is time and cost
consuming. Therefore the usage of experimental techniques is still essential. Here, two
methods can be used: wind tunnel and free flight experiments.
   But both methods have disadvantages. Because of the mounting of the models in
wind tunnel tests there are interferences of the flow. Furthermore there are no regards
to coupling effects as well as interactions that would arise by the enabling of all six
degrees of freedom. The adversarial of free flight techniques are the non-reproducible
conditions (atmospheric disturbances), the high costs and also the risks of manned
flight tests.
   The goal of the Chair of Flight Dynamics is to improve the experimental techniques
conducted so far and thus to design and develop a method and algorithms for
determining the aerodynamic and flight mechanical parameters by wind tunnel free
flight tests. The method should be adaptable on a multiplicity of wind tunnels and
aircraft. The project will be completed by a validation of the method in the wind tunnel
of the Chair of Flight Dynamic with an aircraft with variable static longitudinal
stability.

                 2 HARDWARE CONCEPTION AND MODELING

2.1 Position and Attitude Detection
   This determination can be done by using sensors inside the aircraft, as within
“classic” aircraft identification and control, or outside. As the aircraft only moves in a
small area, referred to the geodetic system, the use of sensors outside the model is
easily possible. Therefore, a 3-D camera system will be used which has several
advantages. No sensors have to be integrated into the aircraft and the position and
attitude must not be calculated by integrating other signals. All other required signals,
as e.g. rate of turns or accelerations result from derivation. Hence there is no drift in
the signals.
   As the measuring principle is based on the assignment of points on the aircraft, the
accuracy depends mainly on the distance of the points to each other, thus the size of the
model, and the dimension of the measuring section which has to be detected by the
cameras. Tests showed that a sample rate of much more than 200 Hz with an
achievable best accuracy of 0.3 mm for each point will be possible.
   The data stream is transported to the real time hardware via a network stream.
2.2 Wind Tunnel
   Wind tunnels can be divided by several characteristics. Besides the classical types,
Eiffel- and Göttinger wind tunnels, several special tunnels, e.g. tailspin and shoot
tunnels, exist. Further classification is the type of measuring section, such as the form
of the cross section and whether it has an open or closed test section.
   For wind tunnel free flight test the first criterion for the selection of the wind tunnel
is the model. This prescribes the size of the measuring section and the producible
speed. Besides these parameters the quality of the free stream, preferably low
turbulence and free of vortices, is the main criterion.
   For the validation tests, the Chair of Flight Dynamics owns a low speed wind tunnel
of Göttingen type with a ∅ 1,5 x 3 m test section. The free stream of the wind tunnel
was studied by traversing a five hole and a split film probes.
2.3 Real-Time Hardware
   A real time system ds1103 from dSPACE is chosen, which is connected and
controlled via a standard PC. One major advantage of this system is the possibility to
generate code out of MATLAB and Simulink.
   A graphical user interface will monitor the process and give the user the ability to
influence and control the experiment.
2.4 Aircraft and their components
   Two aircraft were developed for this project. One is a flying wing and the other a
standard configuration. The requirements were mainly given by the characteristics of
the existing wind tunnel.
Several tools were used, numeric ones like DATCOM, XFLR5 and Vorlax, as well
as data for the components from the test benches and a generic six degree of freedom
simulation to analyse the Eigen values, command actions and disturbance reactions.
The longitudinal stability can be varied by adjusting a sinker in the fuselage of the
airplane. The numerical data were validated by static wind tunnel force and moment
tests.
For the choice of the components, several automated test benches are available as well
as databases, which were created with them. These are a servo, an accumulator and an
actuator test bench. They can measure the static and dynamic behavior of the
components, preprocess the results and provide models of the components in
MATLAB and Simulink. For example, the behavior of a servo or an engine is
simulated as a variable PT2 element with reaction time.
As the transmission time of the control commands of standard HF-Links is in the range
of 25 ms, which will be too high for a automatic control system, a custom-made small
and lightweight HF-installation with a transmission time of 11 ms was built.
An on-board computer, which was developed for this project, collects data from
potentiometer on the servos, an angular velocity sensor and the battery power and
sends them via telemetry to the real time hardware in the control room.

                                 3 IDENTIFICATION
   The goal of the identification is to describe the dynamics of a physical system by a
mathematical model. Hence the transfer functions between the input and output signals
are wanted. The procedure of acquiring them is the same for all identification
algorithms. In free flight experiments, the reactions of the model to different input
maneuvers are recorded. Then, the mathematical model is subjected to the same inputs.
   The real results are compared to those of the model and the parameters are tuned
until the discrepancies between both systems are as small as possible.
   Consequently, the identification consists of the following steps. First, an
identification algorithm has to be selected. A model describing the behavior of the
airplane has to be chosen. A cost function compares the differences between the data
and tunes the parameter. This can be done in the time as well as the frequency domain.
Finally, the maneuvers have to be designed.
3.1 Adaptive Online Parameter Identification
   The first aim of the adaptive online parameter identification is, as the name implies,
to adapt the identification maneuvers autonomously to fit best for the Eigen values of
the aircraft to achieve optimal results. The adaptive parameter identification presented
in the following was developed in [1]. The specifications for the development where
the following:
   - The automatism should simulate a wide class of aircraft
   - The automatism should only need little a priori knowledge about the aircraft.
       This implies that the results are independent of the initial values, if they where
       chosen reasonable.
   - The identification should be robust against signal noise.
3.2 Modeling
   The approach pursued here makes use of physical insight into the system to define
the model, and limits the identification process to the estimation of the model
parameters, which in most cases have a definite physical meaning. The model which is
used in this project consist of the equations of motion of a rigid body in six degrees of
freedom, which are well defined in standard literature, e.g. [2].
   Because of the restricted processing power, a linear model was chosen. Other
authors already reported good results with such a model, e.g. [3], [4] and [5].
Linearisation has several advantages, e.g. simpler and faster computation because of
lower complexity, division into longitudinal and lateral motion, usage of efficient
matrix operations in MATLAB and easier automatic adaption because the methods of
the linear system theory (e.g. Eigen value decomposition) can be used. The main
disadvantage is the limitation of validity to a small range around the trim point. This
means that the flight envelope has to be split into several trim points, control
deflections have to be kept small, the model has to be trimmed and the trim values
must be captured.
3.3 Identification Algorithm
   Various types of identification algorithms have been successfully used in aircraft
parameter identification. A complete survey is given in [6]. The most common online
algorithms are the Recursive Least Squares (RLS) algorithm and its varieties, which
were implied in this project.
   The solution is attained by solving a linear system of equations. The recursive least
squares algorithm does the same in every time step when new measurements are
available, but does not need to solve the linear system again and instead calculates
updates to the solutions of the time step before.
   The variety of the RLS which has been chosen here is the so-called Fourier
Transform Regression. This algorithm makes use of the fact that the Fourier
transformation is a linear one, and thus the model parameters do not change, when
inputs and outputs are transformed into the frequency domain:
   x& = Ax + Bu o → • jω~      x = A~
                                    x + Bu~ (Tilde denotes values in the frequency domain). (1)
   The discrete variety of the Fourier transformation is the z-transformation, which
may be calculated recursively, thus being ideally suited for online implementation. The
transformation is calculated by
   ~
   x k = e − jϖ∆t x k + ~
                        x k −1                                                              (2)
   where indices denote time steps. The frequency vector ϖ , consisting of a set of
discrete frequencies of interest, can be selected according to the needs of the
application, which in this case means that it covers the rigid body dynamics of the
vehicle in question. Leaving out the zero frequency eliminates constant deviations like
biases and removes the influence of trim values. The same goes for the very low
frequencies, which contain only sensor drifts or similar. On the other end of the
spectrum, sensor noise is eliminated by setting the maximum frequency appropriately.
Finally, the computational burden can be varied with the number of frequencies to be
considered.
For the stimulation of the airplane’s motion the following sequences were used
which differ mostly in the frequency spectrum. The dutch roll is stimulated via a
rudder doublet. Aileron doublets are used for the roll imagination. The short period is
identified via an elevator 1123-input. The manoeuvres are adapted in amplitude and
duration.
3.5 Offline Identification
   As the online identification is restricted in the processing power and time, offline
parameter identification will be done after the tests. This has the advantage that e.g.
more complex (e.g. nonlinear) models can be built and the original flight data can be
preprocessed.
   The offline identification consists of a Flight Path Reconstruction (FPR) and an
equation error method. Because of the measuring concept, sensor biases and drifts are
not to be expected, so that an equation error method is sufficient. The FPR is used for
evaluation of the flow angles, as the models will not be equipped with such sensors.
For a detailed description of the two methods, the reader should refer to [6].

                                     4 CONTROL
   The control algorithm is used to keep the aircraft inside the free stream and to
position it for the manoeuvres. It is switched off during the identification manoeuvres,
but the reaction of the model is monitored to switch on the controller if the airplane
starts to leave the free stream.
   The algorithm has to be usable for several aircraft and adaptable, because the a
priori knowledge about the aircraft characteristics should be kept low. Besides, the
aircraft can be very agile and exposed to high frequency disturbances.
   Because of these requirements an explicit method is chosen. Several concepts were
studied, but because of the high non-linearity of the aircraft the non-linear dynamic
inversion is chosen. As the robustness against uncertainties in the parameters and
model data is quite low, it is expanded via an adaptive element, consisting of a neural
network. Simulation studies [7] as well as applications [8] have shown that this
approach maintains stable performance under large variations in the aircraft and
environment.
   To avoid problems with non-linear rate and deflection saturations, which could
destabilize the system and to invert the actuator dynamics, the algorithm is additionally
expanded with a Pseudo Control Hedging (PCH) algorithm.
4.1 Non-linear Dynamic Inversion
   The goal of the non-linear dynamic inversion is to find a non-linear state
transformation so that the resulting system has a linear input / output behavior. Hence
every output only depends on one pseudo-control. For a full explanation the reader
should be referred to [9].
   The theory is quite complex and since now there is no standardized method for
stability analysis. Therefore, to proof stability, the way of simulation the process is
chosen.
To bypass singularity problems caused by an ineffective control matrix the inversion
is splitted into time scale regions. The structure refers to [10], in which an
apportionment of the dynamics of the aircraft in three layers, namely the rotation,
attitude and course dynamics is done. Arranged behind, a distance controller,
consisting of a PI-Element, converts the commanded positions into the course
dynamics block. The structure of the nonlinear dynamic inversion is shown in Fig. 1.

                  Fig. 1 Structure of the nonlinear dynamic inversion

In every layer, the dynamic systems are substituted via a linear system of first order.
   The basis of the rotation dynamics and inversion is the law of conservation of
angular momentum, which calculates the desired moments. But the conclusion to the
control movement is not available in analytical form. Holzapfel [10] proposes a local,
approximate inversion. The actual momentums are subtracted from the commanded,
which has the effect of the natural damping of the aircraft. Hence there is only a linear
coherency across an integrator in the rotation dynamics.
   The inversion of the principle of linear momentum is the basis of the attitude
dynamics inversion. The question arose why not to use the Euler angles instead of the
aerodynamic angles, as they can be measured directly whereas the aerodynamic angles
only can be approximated. But by using the Euler angles the crucial aerodynamic
angles would become an internal, unobservable dynamics. The formulas where
developed under the effect of wind and turbulences.
   The basis of the course dynamic is the point mass differential equation in the wind
fixed coordinate system. At this the transverse force Q and the sideslip angle, as it is
always commanded to zero, is neglected.
4.2 Pseudo-Control Hedging
   Because of the high dynamics of the aircrafts the actuators have to be taken into
account. Furthermore, the controller could command unrealizable values for which the
output value could not follow its command.
   One solution would be to invert the actuators but this would be very complex and
would enhance the order of the system. Johnson [11], [12] uses another approach,
called Pseudo Control Hedging, to evaluate the difference between the expected and
real process reaction. This is done by measuring the actuator positions and using this as
input into the reference model of the aircraft. The approximated dynamic is slowed
down by this difference. Through this the actuator is taken into account and moving
into saturations will be avoided.
   The disadvantage of the PCH is that it is no pure feed forward control any more and
the reference model has to taken into account during the stability analysis.
4.3 Adaption via Neural Networks
As mentioned before, the function of the neural network is to compensate uncertainties
in the parameters and model data and hence stabilize the control algorithm.
Single Hidden Layer (SHL) Perceptron Neural Networks are used, which are universal
approximations for any smooth nonlinear function [13]. For the update, a back
propagation algorithm is used.
The work is still in progress. Current investigations are running, regarding in which
layer of the dynamic inversion a neural network is useful, which input/outputs should
be used, how many neurons should be used and tests with different learning strategies.

                            5 PRELIMINARY RESULTS
The whole process is being simulated in the MATLAB/Simulink environment. The
simulation is generic and adaptable. New aircraft, sensor, engine or environment
models can be loaded via standard data formats. As the real time hardware is coded via
MATLAB/Simulink, the whole process is ported onto the hardware without costs and
sources of error.
    Three aircraft were tested, the two aircraft realised for the project and the Dornier
Do228. This should guarantee the independence and adaptive ability of the algorithms.
Also several tests with different environments took place.
    After the start, the identification algorithm begins with the initial values which
should accomplish marginal maneuvers and are given by the user, analyzes these
maneuvers and adapts them. In the meantime, the controller brings the aircraft back to
the starting point of the maneuver. This loop is repeated till the maneuver fits to the
Eigen motion: Then the identification process will be repeated three times with this
maneuver for statistically firm data.
    The results of the tests showed that the adaptive identification algorithm only needs
little previous knowledge about the system and is insensitive compared to different
initial values. As well the control algorithm, even with the missing neural network,
shows good behavior. Further tests, concerning the robustness against uncertainties
have to be accomplished.

                            6 CONCLUDING REMARKS
   The two experimental techniques free flight and wind tunnel experiments for the
parameter identification are merged together and expanded to a free-flight wind tunnel
test technique. This method could provide the chance to create more authentic flight
mechanical and aerodynamic parameters of the highly coupled system aircraft.
   An experimental validation environment, consisting of a 3D-Camera-System for
position and attitude detection, a real time hardware and aircrafts was composed. The
hardware could be modeled by several numerical tools as well as test beds.
   An online identification algorithm, based on a Fourier transformation regression,
adapts the maneuvers to fit best for the aircraft. A separate offline identification
algorithm, based on a Flight Path Reconstruction and an equation error method is used
which should bring better results.
   A non-linear adaptive controller, consisting of a nonlinear dynamic inversion,
Pseudo Control Hedging and a Neural Network keeps the aircraft inside the free stream
and is used to position the aircraft after an identification maneuver.
Even though due to the missing NN implementation there were no free flight tests
yet, the work done so far gives confidence for the realisation of the test technique.
Final tests in the wind tunnel of the Chair of flight dynamic will be accomplished in
the first quarter of 2009.

                                  7 REFERENCES
[1]    Wolf, C.: Adaptive Parameteridentifizierung in Echtzeit. Diploma Thesis, Chair
       of Flight Dynamics, RWTH Aachen University
[2]    Brockhaus, R.: Flugregelung. 2001, Springer
[3]    Morelli, E. A. – Klein, V.: Application of System Identification to Aircraft at
       NASA Langley Research Center. Journal of Aircraft, Vol. 42, 2005. (p.12-25)
[4]    Morelli, E.A.: In-flight System Identification. AIAA-98-4261, 1998
[5]    Rusnak, I. – Guez A. – Bar-Kana, I.: On-Line Identification and Control of
       Linearized Aircraft Dynamics. IEEE AES Magazine, Vol. 6, 1992. (p. 56-60)
[6]    Jategnonkar, R. V.: Flight Vehicle System Identification: A Time Domain
       Methodology. American Institute of Aeronautics and Astronautics, Inc., 2005
[7]    Calise, A. – Lee, S. - Sharma, M.: Development of a reconfigurable flight
       control law for tailless aircraft. AIAA Journal of Guidance and Control, Vol.24,
       No.5, 2001. (p. 896-902)
[8]    Brinker, J. – Wise, K.: Flight testing of a reconfigurable flight control law on
       the X-36 tailless fighter aircraft. AIAA Journal of Guidance, Control and
       Dynamics, Vol. 24, No. 5. (p. 903-909)
[9]    Khalil, H. K.: Nonlinear Systems. Prentice-Hall Advanced Reference Series
       (Engineering), 2002
[10]   Holzapfel, Florian: Nichtlineare adaptive Regelung eines unbemannten
       Fluggerätes. Phd Thesis, Chair of flight mechanics and flight guidance, TU
       Munich
[11]   Johnson, E. N. – Calise, A.J.: Pseudo Control Hedging: A new Method for
       Adaptive Control. Advances in Navigation Guidance and Control Technology
       Workshop, Alabama, 2000.
[12]   Johnson, E. N. – Calise, A.J.: A Six Degree-Of-Freedom Adaptive Flight
       Control Architecture for Trajectory Following. AIAA 2002-4776, AIAA
       Guidance, Navigation, and Control Conference and Exhibit, Monterey,
       California, 2002
[13]   Lewis, F.L. – Jagannathan, S. – Yesildirek, A.: Neural Network Control of
       Robot Manipulators and Nonlinear Systems. Taylor & Francis Ltd., 1999
You can also read